Protection against Cloning for Deep Learning

Abstract

The susceptibility of deep learning to adversarial attack can be understood in the framework of the Renormalisation Group (RG) and the vulnerability of a specific network may be diagnosed provided the weights in each layer are known. An adversary with access to the inputs and outputs could train a second network to clone these weights and, having identified a weakness, use them to compute the perturbation of the input data which exploits it. However, the RG framework also provides a means to poison the outputs of the network imperceptibly, without affecting their legitimate use, so as to prevent such cloning of its weights and thereby foil the generation of adversarial data.